Inspiration

Brainstorming is often the easiest part of a team project. The real challenge comes afterward: organizing ideas, exploring potential solutions, and figuring out where to start. We noticed that even when teams have great ideas, those ideas can easily get lost in scattered notes and conversations. Moving from an open-ended discussion to a clear direction often takes more time and effort than expected. We wanted to make that transition easier, so teams could spend less time figuring out what to build and more time actually building it.

What it does

Project Room — from ideas to action.

Project Room is an AI-powered collaborative workspace that helps teams turn early-stage ideas into structured plans and actionable tasks.

Teams can brainstorm freely with AI, which helps organize discussions, research the problem space, analyze existing competitors, and identify challenges worth solving.

Based on those insights, Project Room helps teams refine their direction, develop a clearer strategy, and break it down into concrete tasks that can be managed collaboratively. Less chaos, more clarity.

How we built it

Project Room is a TypeScript monorepo with:

React 19 and Vite for the web application.
Express 5 for the API.
Embedded SurrealDB for project data and realtime updates.
Clerk Organizations for authentication and team workspaces.
LLM models for structured AI output.
Nebius for frequent AI operations.
Groq Whisper and browser speech recognition for meeting transcription.
SurrealDB Agent Memory for retrieving information from uploaded project files.
Tavily for grounded brainstorming research.
WebSockets, WebRTC, Zod, and Vitest.

AI output is validated against project data before it is stored. Suggestions are kept separate from approved records, ensuring that a model cannot directly create official tasks, ideas, assignments, or polls.

Challenges we ran into

The hardest challenge was making AI useful without letting it silently change the team’s work. We created a structured approval workflow and validated generated IDs, quotations, sources, and project boundaries.

Realtime collaboration and meetings also required coordinating SurrealDB live queries, WebSockets, WebRTC signaling, transcription, and AI processing. Another challenge was maintaining useful context without sending the entire project to the model, so we combined recent conversations, project-scoped search, meeting transcripts, and retrieved file passages.

Accomplishments that we're proud of

Built a connected workflow from conversation to approved task.
Created an AI meeting room with transcription, summaries, polls, and follow-up suggestions.
Added grounded answers across chats, files, tasks, polls, and meeting transcripts.
Enforced project isolation and human approval throughout the AI workflow.
Built realtime collaboration across multiple browser sessions.
Developed structured brainstorming with research, strategies, ideas, and explicit approval.
Created an evaluation runner that measures action-item precision, recall, F1, and assignee accuracy.

What we learned

We learned that trustworthy AI collaboration depends less on generating more content and more on provenance, validation, and clear control boundaries. Showing why a suggestion was made and requiring confirmation before execution made the AI feel like a teammate instead of an unpredictable automation.

We also learned that project context works best when recent conversation, structured records, search results, and durable document memory are treated as different kinds of evidence.